AI Law, Risk & Governance

Sovereign AI: European Models and Cloud Offerings

Marin Niehues7 April 202511 min read

Hand signing a paper

Sovereign AI means that your company controls the operation and the data of its AI and can change provider; European models and clouds are the obvious option for this. The question is concrete for the executive team, because AI applications are moving into core processes and creating dependencies that can later be undone only at high cost. This article sorts the offerings by decision criteria and describes what an organization must deliver for sovereignty to reach daily operations. This article provides an overview and does not constitute legal advice.

What does sovereign AI mean for companies?

Sovereign AI means that a company determines for every AI application where the data resides, who is accountable for operation and how quickly the model or provider can be replaced. In decision papers, the term is used in very different ways. For assessment, it helps to break it down into four dimensions that can be tested separately:

  • Data sovereignty: where do inputs and outputs, including logs, reside, and who manages the keys?
  • Operational sovereignty: who runs the platform, who can change or shut it down, and who responds to incidents?
  • Model choice: can the model be replaced without rebuilding the application?
  • Ability to switch: is there a tested route to a second provider?

The dimensions are independent of one another. A European model on a non-European platform satisfies model choice, but leaves a gap in operational sovereignty. An open model of American origin in a European data center can satisfy all four dimensions. The origin of the model and the location of operation are therefore separate decisions.

What sovereignty level does an AI application need?

The sovereignty level of an AI application follows from the class of data it processes: the more confidential or regulated the data, the higher the level. The classification applies to each application individually and never across the whole company.

Most companies have applications at all three levels. An assistant for marketing copy works at level 1, an analysis of sales data at level 2, an application with health or personnel data at level 3. Raising everything to level 3 means paying for protection that many applications do not need and giving up functionality. Leaving everything at level 1 means taking a risk with sensitive data for which executive management is accountable.

Which European models are there?

The best-known European model provider is Mistral AI from France, which offers its own language models, some with open weights. Mistral sells further models commercially through its own platform and offers its own assistant, Le Chat. In addition, European platforms run open models in EU data centers, for example at STACKIT or IONOS.

Open models of American or Asian origin widen the choice. In European or own infrastructure, they meet the requirements for data and operational sovereignty, provided their license permits use in the EU. Models under Apache 2.0 or MIT can be used freely, while Llama licenses contain restrictions, for some multimodal releases also for companies based in the EU. The trade-off between open and commercial models is covered in the article Open-source models or commercial models.

At European level, funding initiatives such as OpenEuroLLM for open European language models and the planned AI Gigafactories for computing capacity add to this. In OpenEuroLLM, around 20 research institutions and companies are developing open models for all official EU languages, funded by the Digital Europe Program. They will improve the offering over the medium term. For decisions in the current portfolio, what counts are the models and platforms available in operation today.

Four criteria apply to the choice of a model: performance on your tasks, the license terms, operating effort and the maturity of the provider. You measure performance on your own use case; public leaderboards offer only orientation. A comparison on a sample of real documents from the business function provides a sounder basis within a few days than any leaderboard. The article Large language models compared gives a market overview.

Is Mistral better than ChatGPT?

For companies with high data sovereignty requirements, Mistral is often the better choice, while the strongest American models lead at the performance frontier for open, demanding tasks. The question can therefore be answered per application, and the answer comes from an evaluation with your own cases.

In favor of Mistral are the provider's European base, open weights for part of its models and thus the option of running the model yourself or with a European cloud provider. In favor of ChatGPT and the OpenAI models are the range of functions and performance on complex tasks. A company that connects both through an abstraction layer need not commit and chooses, per application, the model that passes the test.

What is a sovereign cloud?

A sovereign cloud is a cloud offering whose operation, staff and key management sit with a European company, separate from the infrastructure of the global parent group. The large American providers are responding with it to European customers' requirements for operational sovereignty.

The offerings follow two patterns. In the first, the hyperscaler itself runs a separate European cloud with a European entity and EU-based staff, as AWS has announced with its European Sovereign Cloud. In the second, a European company licenses the technology and runs it itself. Examples are Delos Cloud, an SAP subsidiary that provides Microsoft technology for the German public sector, and the T-Systems Sovereign Cloud based on Google Cloud.

The sovereign variants shift operation and keys to European entities, while dependence on the technology provider remains. Updates, the range of functions and prices are still set by the manufacturer. The legal assessment, for instance of access by US authorities under the CLOUD Act, is a matter for your legal department. The article AI and data transfers to US providers describes the framework.

Cloud offerings: where the models run

There are three routes for operation:

  • Hyperscaler with a European region or its own sovereign cloud: a broad range of functions with low operating effort, but dependence on proprietary services.
  • European cloud providers: these include STACKIT and IONOS from Germany and OVHcloud from France. The range of functions is more focused, and control over contract and operation is tighter.
  • Own operation in your own data center or in rented space: the highest control with the highest effort for operation and specialist staff.

A European region secures the storage location. Access rights and key management as well as the contract terms are decided by the overall set-up.

What is the BSI C5 attestation?

The BSI C5 attestation is the report of an independent auditor confirming that a cloud service meets the requirements of the C5 criteria catalog of the German Federal Office for Information Security (BSI). C5 stands for Cloud Computing Compliance Criteria Catalog. The catalog sets out minimum requirements for information security, supplemented by information on data location, jurisdiction and disclosure obligations towards authorities.

There are two types. A Type 1 attestation confirms the suitability of the controls at a reference date. A Type 2 attestation additionally confirms their effectiveness over an audit period and therefore carries more weight. In the German federal administration and the German healthcare sector, C5 is required for cloud services, and many companies use it as a selection criterion in procurement. For your review, what counts is the attestation of the specific service the AI application uses, and its type.

A decision framework in four steps

Delivery practice suggests a sequence that exposes wrong decisions early. First, classify the data an application is to process. Second, assign a sovereignty level to each planned application. Third, build an abstraction layer: the application talks to the model through a uniform interface, so that switching remains a matter of configuration. Fourth, actually carry out the switch before go-live. Only this test reveals which dependencies sit in prompts and data formats. The European Data Act adds support: it is in force, applies in stages and will oblige providers of cloud services to set out the switch to another provider in the contract and prohibits them from creating obstacles to that switch.

What the organization must deliver

In practice, the obstacles usually lie in responsibilities. Procurement chooses the platform, the business function the model, IT the operation, and nobody is accountable for the ability to switch. A product owner with clear decision rights over requirements and platform as well as over the switching route closes this gap. This includes a set of metrics in the AI portfolio that makes dependencies visible, for example the share of applications with a tested second provider. The article AI management: what executive leadership must steer shows how this steering is built as a whole.

Niehues Consulting as a partner for sovereign AI initiatives

Niehues Consulting supports such initiatives according to the Niehues Delivery Model in the phases Understand, Build and Implement. In the Understand phase, business analysis and requirements as well as the business case are developed, from which the sovereignty level per application is derived. In the Build phase, prototypes follow that prove the benefit in the real process and contain the model switch as a fixed test case. In the Implement phase, the solution goes into operation with internal or external teams.

Three of the model's assessment questions matter most for sovereign AI. Who decides, who delivers, who uses? Which roles, decision paths and interfaces does operation require? Who is accountable for operation from day one? Founder Marin Niehues works from the delivery perspective: he builds organizations that deliver and brings AI into operation, in projects from VC-funded scale-ups to DAX 40 groups. The entry point is a free 30-minute initial consultation, which you can request here.

Frequently asked questions

Are there European alternatives to ChatGPT?

Yes, Mistral AI offers its own assistant, Le Chat, and models for businesses. In addition, there are open models run by European cloud providers such as STACKIT, IONOS or OVHcloud. Which option fits is decided by a test on your own tasks.

Is a European cloud automatically sovereign?

No, the location of the data center only governs where data is stored. Sovereignty comes from control over keys and operation as well as clear contract terms. A tested switching route completes it.

Can American models be operated in a sovereign way?

Open models of American origin meet the requirements for data and operational sovereignty in European or own infrastructure, provided their license permits use in the EU. Commercial models from American providers achieve greater operational sovereignty through sovereign cloud offerings, while dependence on the manufacturer remains. The contract terms are decisive.

Which European cloud providers offer AI services?

Well-known European cloud providers with AI services include STACKIT and IONOS from Germany and OVHcloud from France. They run open language models in their own data centers and provide computing capacity for your own models. Their range of functions is more focused than that of the hyperscalers.

How much does sovereign AI cost compared with US services?

Sovereign AI usually costs more operating effort per application, because functions that hyperscalers supply ready-made are missing or because own infrastructure has to be run. You limit the extra cost by running only level 2 and level 3 applications in a sovereign set-up. The calculation belongs in the business case for each application.

Does an AI provider need a BSI C5 attestation?

A BSI C5 attestation is required for cloud services in the German federal administration and healthcare sector and is a strong selection criterion for companies. For level 3 applications, you should require a Type 2 attestation for the specific service used. An attestation does not replace your own review of the contract and data flow.

Let’s talk about your initiative.

In 30 minutes we clarify where your organization stands and which step will pay off most now.

Request a consultation

30 minutes · free · no obligation